Description
Beam is an open-weight model developed by Reflection, featuring 501 billion parameters with 23 billion active. It is designed to handle coding, reasoning, and agentic workloads efficiently. Beam's capabilities are a result of significant investments in pretraining and reinforcement learning (RL). The model was pretrained on 23.8 trillion high-quality tokens from diverse sources, ensuring it matches or outperforms similar-sized open models. The RL training involved over 100 million rollouts on 10.5K NVIDIA GB300 GPUs, focusing on high-compute efficiency.
Read More
Beam is competitive with larger models like GLM 5.2 and approaches Qwen 3.8-Max on coding and agentic tasks. It excels in inference efficiency, providing more intelligence per token at a lower cost. The model is undergoing final evaluations, with weights and technical reports to be released soon.
The model's architecture includes a sparse Mixture-of-Experts (MoE) design, allowing for efficient reasoning and coding capabilities. Beam's RL training emphasizes stable learning, even with asynchronous policy gradients, ensuring robust performance across various tasks.
Beam is suitable for enterprises looking to integrate advanced AI capabilities into their workflows. It supports a range of applications, from software development to machine learning workflows, and is adaptable to different environments. Reflection plans to release Beam under an Apache 2.0 license, making it accessible for developers to build upon.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Beam's Core Features
501 billion parameters
Sparse Mixture-of-Experts model
Pretrained on 23.8 trillion tokens
High-compute reinforcement learning
Competitive with larger models
Efficient inference performance
Open-weight model
Apache 2.0 license
How to use Beam?
Sign up: Join the waitlist for early access
Configure: Set up the model for your specific tasks
Use: Integrate Beam into your workflows
Optimize: Adjust reasoning effort for efficiency
Beam's Use Cases
- Enterprise Coding
- Advanced Reasoning
- Agentic Workloads
- Machine Learning Workflows
- Software Development
